wearable-activity-lstm

Wearable Activity Classifier – Group ___

Task

Classify a 100-step, one-feature sensor sequence into Stationary, Walking, or Running.

Model selected

• Architecture: LSTM
• Input shape: (100, 1)
• Output classes: 3
• Parameters: 4,451

Training data

Synthetic signals generated in the class notebook. The dataset was designed for teaching and is not a real wearable benchmark.

Evaluation

• Test accuracy: 0.913
• Training time in our run: 5.4 seconds

Why we selected this model

LSTM gave the second-highest accuracy (91.3%) among all four models while using far fewer parameters than the CNN (4,451 vs 49,475). It also clearly outperformed SimpleRNN (46.7%), showing that LSTM's gated memory handles this 100-step sequence much better than a basic RNN. We prioritized this balance of accuracy and efficiency over the CNN's marginally higher (100%) accuracy.

Limitations

• Synthetic, simplified data
• One sensor feature only
• No testing across real users/devices
• Not intended for health, safety, or production use

Team learning note

We learned that a more complex architecture doesn't always win — SimpleRNN struggled with this 100-step sequence, but LSTM's gating mechanism handled it far better with only a small increase in parameters over the RNN.